[FA2] Use kernels-community/aiter-flash-attn-ck Hub kernel for ROCm FlashAttention (drop aiter dependency)

[Abdennacer-Badaoui]

What

Replaces the aiter pip-package attention backend with the Hub kernel kernels-community/aiter-flash-attn-ck, loaded via kernels (same pattern as flash_hub / sage_hub).

• Backend aiter -> aiter_fa2_hub; no more aiter pip dependency.
• The -ck CK build is bf16-only, so the dtype constraint is bf16 (fp16 is rejected).

Validation (AMD MI300, ROCm 7.2, torch 2.11)

Numerical parity vs PyTorch SDPA (random q/k/v, B=2, S=1024, H=8, D=64):

dtype	causal	max_abs	mean_rel
bf16	no	3.9e-03	1.3e-02
bf16	yes	3.1e-02	1.4e-02

End-to-end: FLUX.1-schnell, same seed, 4 steps, bf16, native (SDPA) vs aiter_fa2_hub - perceptually identical (mean pixel diff ~= 1.3%).

Prompt: "A photorealistic close-up of a red fox sitting in a snowy pine forest at golden hour, soft rim light, shallow depth of field, highly detailed fur"

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[sayakpaul]

Shouldn't _check_qkv_dtype_bf16_or_fp16 already tackle it?

[Abdennacer-Badaoui]

Not for this case; that helper permits fp16 (it only rejects dtypes outside {bf16, fp16}), so it would let fp16 tensors through to the kernel, which is exactly the input that fails.

The build only ships bf16 (for now) mha_fwd instances. we can try to add fp16 and the backward afterwards

[sayakpaul]

Shouldn't it be aiter attention?

[Abdennacer-Badaoui]

This is inside _native_flash_attention, so it should say "native flash attention"; the original "aiter attention" here was a copy-paste bug, so I fixed it while touching this file.